# How George Hotz actually uses coding agents

tinygrad, comma.ai

**If you take one thing:** `(kimi) in the PR title`, names the model per PR. his tinygrad PRs carry the name of the model that helped, a public log of which model wrote what.

- **Agent:** OpenCode ([source, 2026-04-22](https://geohot.github.io/blog/jekyll/update/2026/04/22/ai-has-no-moat.html))
- **Model:** Kimi ([source, 2026-09-27](https://github.com/tinygrad/tinygrad/commit/9f6d3d882cfe0b8d5e98255b85d9c96dac2966b3))
- **Model:** Claude Opus ([source, 2026-10-11](https://github.com/tinygrad/tinygrad/pull/18733))
- **Review:** Reads and understands each line ([source, 2026-05-24](https://geohot.github.io/blog/jekyll/update/2026/05/24/the-eternal-sloptember.html))
- **Harness:** opencode on a Linux box running local GLM-5.2 ([source, 2026-07-12](https://geohot.github.io/blog/jekyll/update/2026/07/12/i-love-llms.html))
- **Rules file:** tinygrad AGENTS.md (8 lines) ([source, 2026-09-05](https://github.com/tinygrad/tinygrad/blob/479e077ecfcf14d0cf6ea5b500c9d616db812b15/AGENTS.md#L7-L8))

## In their words

> In my opinion, opencode is legit the best coding agent. Not open source cope, like it’s actually number 1.

> variables are bound params (kimi) (#18444)

> Had opus and astra go back and forth on this

> I haven’t seen anyone of them move to a model where they don’t carefully read and understand each line, except in some confined domains.

> I set up a Linux box with opencode on my local GLM-5.2 last week


Archetype (our label): Skeptic. Newest source: 2026-10-11.
Card: https://www.prehistoriq.com/stacks/geohot/ · Data: https://github.com/prehistoriq/stacks/blob/main/data/stacks/geohot.json